CN108985360B - Hyperspectral classification method based on extended morphology and active learning - Google Patents

Abstract

The invention discloses a hyperspectral image classification method based on extended morphology and active learning, which solves the problem that the prior art cannot fully mine the spatial information of the hyperspectral image, resulting in low classification accuracy. The steps are: 1) input hyperspectral image data; 2) reduce dimensionality of the data, extract spectral features, and obtain spatial features through morphological profile transformation; 3) fuse spatial spectral features to divide training and test sample sets; 4) Use the training sample set for SVM classification; 5) Active learning loop, select sample labels by MCLU criteria and AP clustering, and update the training and testing sample sets; 6) Use the new training sample set for SVM classification until the number of training samples reaches the predetermined number. Stop when the number is set to get the final classification result. The invention combines the morphological features of multiple structural elements with active learning, makes full use of the spatial spectrum information, and improves the classification accuracy under the premise of small samples.

Description

Hyperspectral classification method based on extended morphology and active learning

technical field

The invention belongs to the technical field of image processing, and further relates to the technical field of hyperspectral image classification, in particular to a hyperspectral classification method based on extended morphology and active learning. It is used for classification of features in resource exploration, forest cover and disaster monitoring.

Background technique

A hyperspectral sensor, that is, a spectrometer, can simultaneously image a specific area in dozens or hundreds of consecutive bands, and the obtained image is a hyperspectral image. Since hyperspectral imaging involves different wavelength bands, hyperspectral images can obtain rich spectral information, which creates favorable conditions for object recognition and target detection. In recent years, hyperspectral images have been widely used in fine mineral identification, vegetation type identification and classification, urban feature distinction, detection of dangerous environmental factors and disaster monitoring. Due to the hugeness and complexity of hyperspectral data, it is very time-consuming and laborious to label each pixel in the image manually. Therefore, the classification technology of hyperspectral images has become a very important part of hyperspectral image processing technology.

In their paper "A Spectral-Spatial Multicriteria ActiveLearning Technique for Hyperspectral Image Classification" (IEEE Journal of Selected Topics in Applied Earth Observations & Remote Sensing, 2017), S.Patra et al. proposed a hyperspectral method based on active learning and genetic algorithm. Image classification methods. The steps of the method are: 1. Perform PCA dimension reduction on the spectral information data; 2. Perform morphological profile transformation on the spectral information data after dimension reduction with two structuring element sizes to obtain spatial features; 3. Combine the spatial features with the spectral features 4. Through the combination of active learning and genetic algorithm, iteratively carry out SVM supervised classification. It uses two-scale structuring elements to extract the extended morphological contours of the image. However, single-scale or two-scale structuring elements cannot fully mine the spatial information of hyperspectral images, so a satisfactory classification accuracy cannot be obtained. The combination of active learning to select the samples that need to be marked will take too long to calculate the fitness of the samples in each generation of the population, which will lead to too slow to select the marked samples.

Xidian University disclosed a kind of hyperspectral image based on active learning in its patent document "A method of hyperspectral image classification based on active learning" (application number: CN 201410066856.9, application publication number: CN 103839078 B) Classification. The implementation steps of the method are: 1. Extract spectral and spatial features and fuse them into a feature vector; 2. Randomly divide all samples into test data sets and training data sets, and the training data sets are further randomly divided into labeled data 3. Construct an initial ensemble classifier on the labeled dataset; 4. At each iteration, select a fixed number of unlabeled samples with the highest information content for manual labeling according to the new information content metric 5. Use the final ensemble classifier to make predictions. The disadvantage of this classification method is that the use of single-scale structural elements to extract the extended morphological contour of the image also has the problem that the spatial information of the hyperspectral image cannot be fully exploited, so the satisfactory classification accuracy cannot be obtained; The criterion selects the samples to be marked according to the amount of information, and the calculation process is complex, time-consuming, and requires a large number of marked samples. In real life, the labeling of remote sensing data requires manual operation by experts or field surveys, and the cost is quite high. Therefore, how to use as few labeled samples as possible to obtain the highest possible classification accuracy is very important in remote sensing data classification.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the above-mentioned shortcomings of the prior art, and propose a hyperspectral classification method based on extended morphology and active learning. The invention combines the extended multi-structural element morphological profile feature with an active learning framework, fully mines the spatial information of the hyperspectral image through the extended multi-structural element morphological profile feature, combines the spatial feature with the spectral feature, and fully Using spatial information and spectral information, combined with the small sample characteristics of active learning for classification; at the same time, the combination of neighbor propagation clustering and active learning can achieve high-precision classification results under the premise of small samples.

The steps of the present invention to achieve the above objects are: 1) input hyperspectral image data; 2) reduce the dimension of the data, extract spectral features, and obtain spatial features through morphological profile transformation; 3) fuse spatial spectral features to divide training and testing samples 4) Use the training sample set for SVM classification; 5) Active learning loop, select sample labels by MCLU criteria and AP clustering, and update the training and testing sample sets; 6) Use the new training sample set for SVM classification until training Stop when the number of samples reaches the preset number, and get the final classification result. The invention combines the extended multi-structural element morphological features with active learning, and combines the MCLU criterion with the nearest neighbor propagation clustering algorithm in the active learning to fully mine the spatial information of the hyperspectral image, and under the premise of small samples Improved classification accuracy.

The concrete steps that the present invention realizes the above object are as follows:

(1) Input a hyperspectral image to be classified and its corresponding image data set respectively, the image data set contains spectral information and class labels of the data samples;

(2) Principal component analysis method is used to reduce the dimension of the spectral information of the sample, and the first c principal components PC are extracted, where 3≤c≤15, that is, the spectral characteristics of the hyperspectral image;

(3) Perform morphological profile MP transformation on spectral features to obtain morphological profile EMP, that is, the spatial features of hyperspectral images;

(4) The spectral feature and the spatial feature are connected in series by the method of vector stacking, and the feature set OEMP of the hyperspectral image is obtained, and its dimension is 7c;

(5) According to the category label of the sample, randomly select ρ training samples from each category of samples in the feature set OEMP as the training set T, and the remaining samples are the test set U, where 3≤ρ≤6;

(6) Use the training set T to perform SVM supervised classification;

(7) According to the maximum uncertainty MCLU criterion, the samples in the test set U are arranged in order from small to large according to the size of their corresponding MCLU values;

(8) Select the first m samples in the test set U, where 50≤m≤120, and cluster them according to the neighbor propagation AP clustering algorithm to obtain the category to which each sample belongs, and in each category, select The sample with the smallest MULU value is manually marked;

(9) adding the marked samples to the training sample set T, and removing them from the test sample set at the same time, to generate a new training sample set T' and a test sample set U';

(10) Use the training sample set T' to perform SVM supervised classification to obtain the classification results of hyperspectral images;

(11) Judging whether the number of samples in the training sample set T' reaches a preset number, if so, execute step (12), otherwise, return to step (7);

(12) Construct the final classification map from the classification results, and output the final classification map.

Compared with the prior art, the present invention has the following advantages:

First: because the present invention introduces structural elements with multiple sizes in the extended morphology, selects appropriate size intervals, and connects them with the original principal components in series by the method of vector stacking as a new feature of the sample, thereby greatly improving the performance of the sample. classification accuracy;

Second: because the present invention adopts the method of combining the maximum uncertainty criterion MCLU and AP clustering in the active learning process, first select certain samples through the maximum uncertainty criterion MCLU, and then perform AP clustering on them, and select The samples with the smallest MCLU value in each class are labeled, so that the unlabeled samples selected in each iteration are more representative, so that the classification results with higher accuracy can be obtained in a shorter time under the premise of small samples.

Description of drawings

Fig. 1 is the general flow chart of the present invention;

Fig. 2 is the sub-flow chart of active learning in the present invention;

FIG. 3 is a comparison diagram of the overall classification accuracy of the present invention and the prior art, wherein FIG. 3(a) is a comparison diagram of the overall classification accuracy of the present invention and the prior art on Indiana Pines images, and FIG. 3(b) is the comparison diagram of the present invention and the prior art. Comparison of the overall classification accuracy of the prior art on Pavia_U images.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings.

With reference to accompanying drawing 1, the concrete steps of the present invention are as follows:

Step 1, input data;

Input a hyperspectral image to be classified and its corresponding image dataset, the image dataset contains spectral information and category labels of the data samples;

In the embodiment of the present invention, two hyperspectral images were selected and two experiments were performed. The first image is the Pavia_U hyperspectral image with 103 bands and the class label of the image; the second image is the IndianaPines hyperspectral image with 200 bands and the class label of the image;

Step 2, extract spectral features;

Since the high-dimensional characteristics of the hyperspectral image will bring about problems such as complicated calculation and information redundancy, the present invention adopts the principal component analysis method to perform dimension reduction processing on the spectral information of the hyperspectral image, and extract the first c principal components PC, that is, high Spectral signature of a spectral image, where 3≤c≤15. Here, taking the optimal value of c as 10 as an example, the first 10 principal components PC of the hyperspectral image are obtained;

The specific steps of using the principal component analysis method to reduce the dimension of the spectral information of the sample are as follows:

(2.1) Obtain the spectral matrix X _s of the hyperspectral image according to the spectral information of the sample:

Among them, n is the number of samples, p is the length of the spectral information of the sample, x _np represents the p-th dimension value of the spectral information of the n-th sample; X _s represents the spectral matrix of hyperspectral images of n samples, and each row of X _s Both represent a sample with spectral information length p.

(2.2) Calculate the average value of the i-th dimensional spectral information of the sample by the following formula

Among them, i=1,2,3,...p, represents the summation operation, x _ki represents the i-th dimensional spectral information value of the k-th sample, and 1<k≤n;

(2.3) Calculate the covariance value S _ij of the i-th row and the j-th column of the spectral matrix X _s :

表示样本第j维光谱信息的平均值，x_kj表示第k个样本的第j维光谱信息值，且1＜k≤n；Among them, ∑ represents the summation operation, · represents the multiplication operation of the numerical value and the numerical value, j=1,2,3,...p,

represents the average value of the jth dimension spectral information of the sample, x _kj represents the jth dimension spectral information value of the kth sample, and 1<k≤n;

The covariance matrix S of the spectral matrix X _s is obtained by further calculation:

(2.4) Diagonalize the covariance matrix S according to the following formula:

S*q _u =λ _u ×q _u

Among them, q _u represents the u-th eigenvector of the covariance matrix S, λ _u represents the u-th eigenvalue of the covariance matrix S, * represents the multiplication operation between the matrix and the matrix, and × represents the multiplication operation between the numerical value and the matrix, u=1,2,3,...p;

(2.5) Orthonormalize the eigenvectors;

(2.6) The normalized eigenvectors are arranged according to the corresponding eigenvalues, from large to small, to obtain a feature matrix X _z ;

(2.7) Calculate the spectral characteristic matrix X _a according to the following formula:

X _a =X _z *X _s

Among them, X _a represents the spectral feature matrix, and the first c columns of each row of the spectral feature matrix X _a are defined as the spectral features after dimension reduction of the corresponding hyperspectral image dataset samples, 1≤c≤p;

(2.8) Take the first c columns of each row of X _a , which are the first c principal components PC of the hyperspectral image. Here c is taken as 10, then the first 10 principal components PC of the hyperspectral image are obtained.

Step 3, extracting spatial features;

The first 10 principal components PC obtained in step 2 are subjected to morphological profile MP transformation for each principal component PC _h (h=1, 2, 3, ... 10) respectively. When the dimensions of the structural elements are z, 2z, At 3z, the open section and closed section are calculated for each principal component, and a total of 60 morphological sections EMP can be obtained, that is, spatial features;

The selection of the size and number of the structural elements is very important, and plays a crucial role in the classification accuracy of the hyperspectral image. In the embodiment of the present invention, the Pavia_U hyperspectral image takes the size of the structural elements as 20, 40, and 60; IndianaPines hyperspectral image, taking the size of structuring elements as 5, 10, 15;

The specific steps for obtaining the morphological profile EMP by performing the morphological profile MP on the spectral features are as follows:

(3.1) Find the open section of each principal component in the 10 principal components PC:

表示开运算操作，开剖面是一系列膨胀操作加腐蚀操作的结果，R为结构元素尺寸大小；

表示主成分PC_h的第d个开形态学剖面特征，当d为1、2、3时，结构元素尺寸R分别取到z、2z、3z；Among them, PC _h represents the h-th principal component, and h=1, 2, 3,...

Represents the opening operation, the opening section is the result of a series of expansion operations and corrosion operations, and R is the size of the structural element;

Represents the d-th open morphological profile feature of the principal component PC _h . When d is 1, 2, and 3, the structuring element size R is taken as z, 2z, and 3z, respectively;

(3.2) Find the closed section of each of the 10 principal components PC:

表示闭运算操作，其与开运算相反，是一系列腐蚀操作加膨胀操作的结果，当d为1、2、3时，R为结构元素尺寸大小，分别取到z、2z、3z；OP_γd(PC_h)表示主成分PC_h的第d个闭形态学剖面特征；Among them, PC _h represents the h-th principal component, and h = 1, 2, ... 10; the closed morphological section is obtained by using the structural elements of different sizes to perform the closing operation on the same component,

Represents the closing operation, which is the opposite of the opening operation and is the result of a series of corrosion operations plus expansion operations. When d is 1, 2, and 3, R is the size of the structural element, and z, 2z, and 3z are obtained respectively; OP _γd (PC _h ) represents the d-th closed morphological profile feature of the principal component PC _h ;

(3.3) Calculate the morphological profile MP(PC _h ) of the h-th principal component PC _h :

Taking h=1, 2,...10 in turn, and arranging the morphological profile features calculated by each principal component in order, the morphological profile EMP of 10 principal components PC is obtained:

EMP={MP(PC ₁ ), MP(PC ₂ ), . . . MP(PC ₁₀ )}.

Step 4, fusing empty spectrum features;

The spectral feature and the spatial feature are connected in series by the method of vector stacking, and the feature set OEMP of the hyperspectral image is obtained, that is, OEMP={PC, EMP}, which is 70 dimensions;

Step 5, obtaining a training sample set and a test sample set;

According to the class labels of the samples, randomly select ρ training samples from each class of samples in the feature set OEMP as the training set T, and the rest of the samples are the test set U, where 3≤ρ≤6, where the value of ρ is taken as the training set T. 3 is further described in this embodiment;

In the embodiment of the present invention, the Pavia_U hyperspectral image has 9 types of category labels, and a total of 27 samples are selected as the training samples T; for the Indiana Pines hyperspectral image, there are 16 types of category labels, then a total of 48 samples are selected as the training samples T. training sample T;

Step 6, construct an initial classifier;

Use the training set T and the class labels corresponding to each sample in the training set to perform SVM supervised classification;

Step 7, MCLU criterion;

According to the maximum uncertainty criterion (MCLU criterion), the samples in the test set U are arranged in order from small to large according to the size of their corresponding MCLU values;

The MCLU criteria are:

MCLU is based on the geometric distance of the classification hyperplane. By calculating the distance between the sample and each type of hyperplane, the difference between the first two maximum distances is obtained. The smaller the difference is, the sample can be divided into these two categories. If the reliability is similar, the greater the amount of information contained in the sample, the greater the performance improvement of the classifier after adding it to the training sample set.

Calculate the MCLU value of the sample according to the following formula:

Among them, la represents the number of categories of the sample, r ₁ represents the serial number of the maximum value of the distance between the sample and the classification surface, r ₂ represents the serial number of the second largest value of the distance between the sample and the classification surface, and X ^MCLU represents the MCLU value of the sample x.

Step 8, neighbor propagation clustering, and select the samples that need to be marked;

Select the first m samples in the test set U, where 50≤m≤120, cluster them according to the AP clustering algorithm of neighbor propagation, get the category to which each sample belongs, and select the MULU value in each category The smallest sample is marked by experts, and the value of m is taken as 100 here;

AP clustering is a kind of partition clustering method, which classifies data according to the similarity between data objects. There are two types of messages delivered in AP, attraction and attribution. The attractiveness r _t (l, s) represents the numerical message sent from the data l to the candidate cluster center s, reflecting whether the point s is suitable as the cluster center of l. a _t (l, s) represents the degree of belonging of sample l to sample s in the t-th generation. The attribution degree at (l, _s ) is a numerical message sent from the candidate cluster center s to l, reflecting whether l chooses s as its cluster center. The larger _{r t} (l,s) and at (l,s) are, the more likely point s is to be the cluster center, and the more likely that l belongs to the cluster with s as the cluster center. big. The AP algorithm continuously updates the attractiveness and attribution values of each point through an iterative process until τ high-quality cluster centers are generated, and the remaining data points are assigned to the corresponding clusters.

The above steps are as follows to cluster m samples according to the AP clustering algorithm of neighbor propagation:

(8.1) Initialize the attraction matrix R and the attribution matrix A:

Among them, 1<l≤m, 1<s≤m; t is the number of iterations, and t is initialized to 1; r _t (l, s) represents the attractiveness of sample s to sample l in the t-th generation, a _t (l ,s) represents the degree of belonging of sample l to sample s in the t-th generation;

(8.2) The attractiveness of the updated sample s to the sample l is r _t+1 (l, s):

Among them, at (l, _s ') is the attribution degree of sample l to sample s' in the t-th generation;

(8.3) The attribution of update sample l to sample s is a _t+1 (l, s):

(8.4) Sum the attractiveness and attribution of sample l and sample s to get the objective function f(l,s):

f(l,s)=r _t+1 (l,s)+at ₊₁ (l,s)

Further get the corresponding matrix F of f(l,s):

(8.5) Determine whether the size of each element in F remains unchanged or whether the t value is 1000. If so, obtain each category to which m samples belong; otherwise, add 1 to the t value and return to step (8.2).

Step 9, generate a new training sample set and a test sample set;

Adding the samples to be marked obtained in step 8 to the training sample set T, and removing them from the test sample set at the same time, to generate a new training sample set T' and a test sample set U';

Step 10, construct a classifier;

Use the training sample set T' to perform SVM supervised classification to obtain the classification results of hyperspectral images;

Step 11, determine whether the number of training samples reaches a preset number;

Determine whether the number of samples in the training sample set T' reaches a preset number. In the embodiment of the present invention, the preset number of Pavia_U hyperspectral images is set to 590 samples; for Indiana Pines hyperspectral images, the preset number is set to 290 samples; if so, execute step (12), otherwise, return to step (7);

Step 12, obtain the classification result;

The final classification map is constructed from the classification results, and the final classification map is output.

The effect of the present invention will be further described below in conjunction with simulation experiments.

1. Simulation experimental conditions:

The running environment of the simulation experiment of the present invention is: the processor is Inter Core i3-3210M, the main frequency is 3.2GHz, and the memory is 4GB; the software platform is Windows10 64-bit operating system and Matlab R2017a for simulation test.

2. Simulation experimental data:

The hyperspectral images used in the simulation experiments of the present invention include Indiana hyperspectral images and images of the University of Pavia. Indiana hyperspectral image AVIRIS Indiana Pines is a commonly used data in hyperspectral classification experiments. It is imaged by NASA's Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) on the Indiana Remote Sensing Experiment Area in northwestern Indiana, USA. Acquired in 1992. It contains a mixture of crops, grasslands and forest vegetation, with a total of 16 types of features. The size of the entire image is 145×145 pixels, the spatial resolution is 20m×20m, and 200 bands remain after removing 20 clutter bands. The Pavia_U image, a hyperspectral remote sensing image of the Pavia campus, is a commonly used data in classification experiments, which is acquired by the ROSIS sensor. The ROSISI sensor divides the 0.43-0.86m spectrum into 115 bands with a spatial resolution of 1.3m. The image size of the Pavia campus image is 610×340. After removing the clutter band, there are 103 bands left. The image contains 9 types of information.

3. Simulation experiment content and result analysis:

There are two simulation experiments of the present invention.

The simulation experiment 1 of the present invention verifies the rationality of the selection of the size and number of the structural elements of the present invention. When the present invention extracts spatial features through morphological profile transformation, the selection of the size and number of structuring elements in morphological profile transformation is very important, and plays a crucial role in the classification accuracy of hyperspectral images. selected, and the following comparative tests were performed. When other steps are the same as in the present invention, for the Indiana Pines hyperspectral image, the structural element of Comparative Experiment 1 only takes one size and its value is 5; the structural element of Comparative Experiment 2 takes three sizes and its values are 1, 2, 3; the structural elements of Comparative Experiment 3 take five dimensions and their values are 1, 2, 3, 4, and 5 respectively; the structural elements of the present invention take three dimensions and their values are 5, 10, and 15 respectively; for Pavia_U Hyperspectral images, the structural element of Comparative Experiment 1 has only one size and its value is 20; the structural element of Comparative Experiment 2 has three sizes and its value is 1, 2, and 3; the structural element of Comparative Experiment 3 is taken five The dimensions and their values are 1, 2, 3, 4, and 5, respectively; the structural elements of the present invention take three dimensions and their values are 20, 40, and 60, respectively. The classification results are shown in Tables 1 and 2.

Table 1 Comparison of classification results of different structural elements on Indiana Pines images

Table 2 Comparison of classification results of different structural elements on Pavia_U images

As can be seen from Table 1 and Table 2, the present invention has higher classification accuracy compared with the method of taking only one structural element and the method that the size interval of the structural element is too small. It shows that the selection of the size and number of structural elements is very important, and plays a crucial role in the classification accuracy of hyperspectral images. The present invention introduces structural elements with multiple sizes in the extended morphology, and selects the appropriate size. interval, so that the hyperspectral classification can achieve higher classification accuracy.

The simulation experiment 2 of the present invention compares the method of the present invention and the two classification methods of the prior art. The two methods are the hyperspectral image classification method based on active learning and genetic algorithm proposed by S. Patra et al., and the hyperspectral image classification method based on active learning and recurrent belief propagation proposed by Li Jun et al. The method proposed by S. Patra et al. is to first perform PCA dimensionality reduction on the spectral information data, and then use two structuring element sizes to transform the morphological profile of the dimensionally reduced spectral information data to obtain empty spectral features. Algorithms are combined to iteratively perform SVM supervised classification. The method proposed by Jun Li et al. first combines the spectral information and spatial information of hyperspectral images through cyclic belief propagation, and then uses the spatial spectral information for active learning supervised classification.

Because the sampling methods of each method are different, the same number of training samples cannot be obtained, so each method finally selects a similar number of training samples to ensure fairness. In the method proposed by S. Patra et al., 3 samples of each class in the hyperspectral image are selected as the initial training samples. For Indiana Pines images with 16 classes, 48 initial samples are selected, and 20 training samples are selected for each iteration of active learning sample, iteratively 27 times, a total of 588 training samples are selected; for Pavia_U images with 9 categories, 27 initial samples are selected, and 13 iterations are required, a total of 287 training samples are selected; the support vector machine classifier adopts cross-validation way to set parameters. In the method proposed by Li Jun et al., 5 initial samples are randomly selected for each type of IndianaPines image, 80 initial samples are selected, and 10 training samples are selected for each iteration of active learning, which requires 51 iterations, and a total of 590 training samples are selected Samples; for Pavia_U images, there are 9 categories, 10 initial samples are randomly selected for each category, 90 initial samples are selected, and 10 training samples are selected for each iteration of active learning, and 20 iterations are required, a total of 290 training samples are selected. The present invention randomly selects 3 initial samples for each type of Indiana Pines image, then selects 48 initial samples, and the maximum number of training samples preset by active learning is 590; for Pavia_U images, there are 9 types, and each type randomly selects 3 initial samples , 27 initial samples are selected, and the maximum number of training samples preset by active learning is 290. The parameters of the support vector machine classifier were set by cross-validation, and the simulation experiments were carried out for 10 times.

Tables 3 and 4 show the comparison of the overall classification accuracy (OA) average value, the average classification accuracy (AA) average value, and the Kappa coefficient average value of 10 experiments on two images between the present invention and the prior art. SSMAL represents the hyperspectral image classification method based on active learning and genetic algorithm proposed by S. Patra et al., and MPM-LBP-AL represents the hyperspectral image classification method based on active learning and cyclic belief propagation proposed by Li Jun et al.

Table 3 The comparison table of the classification results of the prior art and the present invention on Indiana Pines images

The comparison table of the classification result of table 4 prior art and the present invention on Pavia_U image

FIG. 3 is a comparison diagram of the overall classification accuracy of the present invention and the prior art, wherein FIG. 3(a) is a comparison diagram of the overall classification accuracy of the present invention and the prior art on Indiana Pines images, and FIG. 3(b) is the comparison diagram of the present invention and the prior art. Comparison of the overall classification accuracy of the prior art on Pavia_U images.

It can be seen from Tables 3 and 4 that in the simulation experiment of Indiana Pines images, the present invention has higher classification accuracy and time efficiency than the method of S. Patra et al. The method of Li Jun et al. Although the present invention has some advantages in classification accuracy, the present invention has obvious advantages in running time, which is 5480 seconds faster than the method of Li Jun et al. In the simulation experiment of Pavia_U image, compared with the method of S.Patra et al., the present invention has advantages in time and the classification accuracy is also greatly improved; compared with the method of Li Jun et al. The average required time is 21757 seconds faster than the method of Li Jun et al., which can achieve high classification accuracy in a very short time. From this, it can be seen that the classification accuracy of the present invention is higher, and the time advantage is more obvious when the image data is larger. It can be seen from the overall classification accuracy comparison diagram in FIG. 3 that the classification accuracy of the present invention is higher because the present invention has more reasonable selection of the size and size of structural elements, and the present invention utilizes MCLU criteria and AP clustering in active learning Combined to select samples for marking, the selected samples are more representative and diverse. The time advantage of the present invention is also obvious. This is because the utilization scheme of the method of Li Jun et al. has a process of processing the image, which is very time-consuming, especially when the image data is large. obvious. And the spatial information utilization method of the present invention is that the morphological profile feature is used as the spatial feature to realize the introduction of spatial information, the operation is simple, and a large part of time is saved. The invention combines extended morphology with active learning, uses MCLU criterion and AP clustering to select samples for marking, shortens the time required for classification, and improves classification accuracy.

The parts of the present invention that are not described in detail belong to the common knowledge of those skilled in the art.

The above descriptions are only a few specific examples of the present invention. Obviously, for those skilled in the art, after understanding the content and principles of the present invention, it is possible to carry out forms and details without departing from the principles and structures of the present invention. Various modifications and changes above, but these modifications and changes based on the idea of the present invention are still within the scope of protection of the claims of the present invention.

Claims (5)

1. a hyperspectral image classification method based on extended morphology and active learning, is characterized in that, comprises the steps: (1) Input a hyperspectral image to be classified and its corresponding image data set respectively, the image data set contains spectral information and class labels of the data samples; (2) Principal component analysis method is used to reduce the dimension of the spectral information of the sample, and the first c principal components PC are extracted, where 3≤c≤15, that is, the spectral characteristics of the hyperspectral image; (3) Perform morphological profile MP transformation on spectral features to obtain morphological profile EMP, that is, the spatial features of hyperspectral images; where morphological profile MP transformation is performed on spectral features, when the structuring element sizes are z, 2z, and 3z, respectively At this time, the open section and the closed section of the c principal components PC are obtained respectively, and a total of 6c morphological sections EMP are obtained. The specific steps are as follows: (3.1) Find the open section of each principal component in the c principal components PC:

Among them, PC _h represents the h-th principal component, and h=1, 2, 3, ... c;

Indicates the opening operation, R is the size of the structuring element;

Represents the d-th open morphological profile feature of the principal component PC _h . When d is 1, 2, and 3, the structuring element size R is taken as z, 2z, and 3z, respectively; (3.2) Find the closed section of each principal component in the c principal components PC respectively:

Among them, PC _h represents the h-th principal component, and h=1,2,...c;

represents the closing operation, R is the size of the structuring element; OP _γd (PC _h ) represents the d-th closed morphological profile feature of the principal component PC _h , when d is 1, 2, and 3, the size R of the structuring element is taken as z, 2z, 3z; (3.3) Calculate the morphological profile MP(PC _h ) of the h-th principal component PC _h :

Take h = 1, 2, ... c in turn to obtain the morphological profile EMP of the c principal components PC: EMP={MP(PC ₁ ), MP(PC ₂ ),...MP(PC _c )}; (4) The spectral feature and the spatial feature are connected in series by the method of vector stacking, and the feature set OEMP of the hyperspectral image is obtained, and its dimension is 7c; (5) According to the category label of the sample, randomly select ρ training samples from each category of samples in the feature set OEMP as the training set T, and the remaining samples are the test set U, where 3≤ρ≤6; (6) Use the training set T to perform SVM supervised classification; (7) According to the maximum uncertainty MCLU criterion, the samples in the test set U are arranged in order from small to large according to the size of their corresponding MCLU values; (8) Select the first m samples in the test set U, where 50≤m≤120, and cluster them according to the neighbor propagation AP clustering algorithm to obtain the category to which each sample belongs, and in each category, select The sample with the smallest MULU value is manually marked; (9) adding the marked samples to the training sample set T, and removing them from the test sample set at the same time, to generate a new training sample set T' and a test sample set U'; (10) Use the training sample set T' to perform SVM supervised classification to obtain the classification results of hyperspectral images; (11) Judging whether the number of samples in the training sample set T' reaches a preset number, if so, execute step (12), otherwise, return to step (7); (12) Construct the final classification map from the classification results, and output the final classification map. 2. method according to claim 1, is characterized in that, in step (2), adopting principal component analysis method to carry out the concrete steps of dimensionality reduction to the spectral information of sample as follows: (2.1) Obtain the spectral matrix X _s of the hyperspectral image according to the spectral information of the sample:

Among them, n is the number of samples, p is the length of the spectral information of the sample, and x _np represents the p-th dimension value of the spectral information of the n-th sample; (2.2) Calculate the average value of the i-th dimensional spectral information of the sample by the following formula

Among them, i=1, 2, 3,...p, ∑ represents the summation operation, x _ki represents the i-th dimensional spectral information value of the k-th sample, and 1<k≤n; (2.3) Calculate the covariance value S _ij of the i-th row and the j-th column of the spectral matrix X _s :

Among them, ∑ represents the summation operation, · represents the multiplication operation of the numerical value and the numerical value, j=1,2,3,...p,

represents the average value of the jth dimension spectral information of the sample, x _kj represents the jth dimension spectral information value of the kth sample, and 1<k≤n; The covariance matrix S of the spectral matrix X _s is obtained by further calculation:

(2.4) Diagonalize the covariance matrix S according to the following formula: S*q _u =λ _u ×q _u Among them, q _u represents the u-th eigenvector of the covariance matrix S, λ _u represents the u-th eigenvalue of the covariance matrix S, * represents the multiplication operation between the matrix and the matrix, and × represents the multiplication operation between the numerical value and the matrix, u=1,2,3,...p; (2.5) Orthonormalize the eigenvectors; (2.6) the normalized eigenvectors are arranged according to the size of the corresponding eigenvalues, from large to small, to obtain a feature matrix X _z ; (2.7) Calculate the spectral characteristic matrix X _a : X _a =X _z *X _s (2.8) Take the first c columns of each row of X _a , which are the first c principal components PC of the hyperspectral image. 3. The method according to claim 2, wherein the optimal value of c in step (2.8) is 10. 4. method according to claim 1, is characterized in that, in step (5), the optimum value of ρ is 3. 5. method according to claim 1 is characterized in that, in step (8), m samples are clustered according to the neighbor propagation AP clustering algorithm, and the steps are as follows: (8.1) Initialize the attraction matrix R and the attribution matrix A:

Among them, 1<l≤m, 1<s≤m, t is the number of iterations, initialized to 1, r _t (l, s) represents the attractiveness of sample s to sample l in the t-th generation, a _t (l, s) represents the degree of belonging of sample l to sample s in the t-th generation; (8.2) The attractiveness of the updated sample s to the sample l is r _t+1 (l, s):

Among them, at (l, _s ') is the attribution degree of sample l to sample s' in the t-th generation; (8.3) The attribution of update sample l to sample s is a _t+1 (l, s):

(8.4) Sum the attractiveness and attribution of sample l and sample s to get the objective function f(l,s): f(l,s)=r _t+1 (l,s)+at ₊₁ (l,s) Further get the corresponding matrix F of f(l,s):

(8.5) Determine whether the size of each element in F remains unchanged or whether the t value is 1000. If so, obtain each category to which m samples belong; otherwise, add 1 to the t value and return to step (8.2).

Images